Structural damage non-contact automatic identification method and system

Abstract

The invention discloses a structural damage non-contact automatic identification method and system, and relates to the technical field of structural health monitoring. The method comprises the following steps: determining an excitation incident angle and a sine excitation signal relative to a signal of a to-be-measured structure; determining the inclination angle of an air coupling receiving sensor of a receiving end according to the excitation incident angle, and amplifying and transmitting the received sine excitation signal to obtain feedback time sequence monitoring data; performing denoising processing on the time sequence monitoring data, and performing identification according to the denoised data and the trained damage degree prediction agent model; the recognition result comprises whether the to-be-detected structure has structural damage or not and the structural damage degree when the structural damage exists; the de-noising processing adopts an ensemble empirical mode decomposition method based on a continuous mean square error criterion. The method can solve the problems that traditional ultrasonic guided wave detection is poor in continuity and the automation degree of damage degree recognition is low, and improves the structural damage recognition and classification accuracy based on air coupling ultrasonic guided waves.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and particularly to a non-contact automatic identification method and system for structural damage. Background Art

[0002] Structural health monitoring technology involves using sensors to obtain response information at key positions of a structure, and adopting signal and mathematical statistics algorithms to extract damage-sensitive characteristics from the sampled information and perform feature analysis to determine the current health status of the structure. As a strategy and process for damage identification and characterization of engineering structures, structural health monitoring has the advantages of real-time monitoring and manpower saving, and has currently received extensive attention and is in rapid development.

[0003] The damage identification method based on vibration response uses changes in conditions such as the stiffness and boundaries of a structure to identify structural damage. However, when there are local minor damages in the structure, this method is not sensitive enough. Ultrasonic body waves are usually used to monitor potential local minor damages inside the structure. This technology is based on the principle of body wave reflection, and its single detection range is small, and the detection effect is greatly affected by the surface state of the structure. The ultrasonic guided wave technology can achieve long-distance detection with a single excitation, but this technology requires pre-applying acoustic couplants such as oil, colloid, water, etc. on the test specimens, which is difficult to meet the requirements of high-efficiency continuous detection, and its application range is limited. For example, porous and water-permeable materials, medical devices that prohibit contamination, and other occasions sensitive to couplants are not suitable for using this detection technology.

[0004] With the increase in the service time of engineering structures and the increasing demand for service toughness, it is imperative to develop non-contact internal damage detection technology for structures and an automatic identification algorithm for damage degree based on received signals. The new air-coupled piezoelectric sensor does not require an acoustic couplant. It can directly emit ultrasonic waves to act on the surface of the structure after propagating through the air, without direct contact with the structure to be detected, and can achieve non-contact continuous and uninterrupted detection, greatly improving the detection efficiency and broadening the application range. However, its signal intensity is smaller than that of the traditional ultrasonic body wave technology, and the received signal intensity is not very ideal, which will affect the effective identification of the internal damage degree of the structure to a certain extent. Summary of the Invention

[0005] The object of the present invention is to provide a non-contact automatic identification method and system for structural damage, which can solve the problems of poor continuity of traditional ultrasonic guided wave detection and low automation degree of damage degree identification.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A non-contact automatic identification method for structural damage includes:

[0008] Determine the excitation incident angle of the signal relative to the structure to be measured and the sinusoidal excitation signal;

[0009] Determine the tilt angle of the air-coupled receiving sensor at the receiving end according to the excitation incident angle, and amplify and transmit the received sine excitation signal to obtain the feedback time-series monitoring data;

[0010] Denoise the time-series monitoring data, and identify it according to the denoised data and the trained damage degree prediction proxy model; the identification result includes whether there is structural damage in the structure to be measured and the degree of structural damage when there is structural damage; the denoising process uses the ensemble empirical mode decomposition method based on the continuous mean square error criterion.

[0011] Optionally, the method for determining the excitation incident angle and the sine excitation signal is as follows:

[0012] Set the acoustic refraction angle to 90°, and determine the excitation incident angle θ according to Snell's law:

[0013]

[0014] In the formula, c 空气 represents the acoustic wave velocity in the air medium, c 结构 represents the Rayleigh wave velocity in the structural medium, and θ represents the excitation incident angle of the air-coupled ultrasonic sensor;

[0015] Select the sine signal modulated by the Hanning window as the excitation signal of the signal generator, which is expressed as:

[0016]

[0017] In the formula, x(t) represents the sine signal modulated by the Hanning window; f c represents the center frequency of the modulation signal; T represents the number of cycles of the modulation signal.

[0018] Optionally, the specific steps of denoising the time-series monitoring data, identifying it according to the denoised data and the trained damage degree prediction proxy model, and obtaining the identification result are as follows:

[0019] Use the ensemble empirical mode decomposition method based on the continuous mean square error criterion to denoise the time-series monitoring data to obtain the denoised data;

[0020] Extract the features of the denoised data to obtain the time-domain and frequency-domain feature values;

[0021] Construct a multi-layer feedforward artificial neural network based on error backpropagation as the damage degree prediction proxy model, optimize the parameters of the damage degree prediction proxy model, and use the trained damage degree prediction proxy model to identify the time-domain and frequency-domain feature values to obtain the identification result.

[0022] Optionally, the time series monitoring data is denoised by using the ensemble empirical mode decomposition method based on the continuous mean square error criterion to obtain the denoised data, which specifically includes:

[0023] Set the original signal in the time series monitoring data as x(t), and perform hierarchical decomposition according to the ensemble empirical mode decomposition method to obtain multiple-order intrinsic mode functions:

[0024] Respectively obtain the local maximum and minimum values of the original signal x(t), perform cubic spline interpolation fitting on the envelopes of the local maximum and minimum values, sum and average the fitting envelopes to obtain the average value m(t), and subtract the average value m(t) from the original signal x(t) to obtain the intermediate signal h(t) = x(t) - m(t);

[0025] Judge whether the intermediate signal h(t) satisfies the two conditions of the intrinsic mode function F(t): (a) The total number of signal extreme points is equal to or differs by one from the total number of zero-crossing points; (b) The average value of the upper envelope line fitted by the local maximum values of the signal and the lower envelope line fitted by the local minimum values should be zero; if h(t) satisfies the two conditions, then h(t) is an intrinsic mode component; otherwise, use the intermediate signal h(t) to replace x(t), and loop through the calculation of the previous step until the recalculated intermediate signal h(t) satisfies the conditions;

[0026] When the first intrinsic mode F 1 (t) is obtained, subtract F 1 (t) from the original signal x(t) to get x 1 (t) = x(t) - F 1 (t), let x 1 (t) be the new original signal, and re-execute the operation steps to obtain F 2 (t), and sequentially obtain each-order intrinsic mode function according to the above steps. When the intrinsic mode function is a monotonic function, terminate the signal mode decomposition; finally, the original signal is decomposed into a series of intrinsic mode functions F i (t) and the residual term R(t), which is expressed as:

[0027]

[0028] In the formula, i represents the order of the intrinsic mode function, and M represents the total number of orders of the decomposed intrinsic mode functions;

[0029] Eliminate the different-scale mode aliasing phenomenon through the set and average steps of the mode decomposition process. The specific steps are as follows:

[0030] Add different white noises to the original signal multiple times to generate multiple different noisy signals to be processed: xj y(t) = x(t) + n j y(t);

[0031] where n j y(t) is the white noise sequence added for the j-th time, j = 1, 2..., and the empirical mode decomposition is performed on different noisy signals respectively to obtain multiple groups of intrinsic mode components F i j y(t), and the components of the corresponding order in different groups of F i j y(t) are added and averaged to obtain the final result F i y(t) of the intrinsic mode component of this order:

[0032]

[0033] where F i j y(t) represents the i-th order intrinsic mode component obtained after decomposing the mixed signal x j y(t) with white noise added for the j-th time, and N represents the total average number;

[0034] By the set and average steps, the mode aliasing caused by noise is eliminated. When N reaches the set threshold, the F i j y(t) sets are combined and averaged to cancel out the white noise components with each other, and the ideal intrinsic mode function F i y(t) is obtained; in the ensemble empirical mode decomposition method, the relationship between the average number and the decomposition error is determined by the following formula:

[0035]

[0036] In the formula, e represents the ensemble empirical mode decomposition error, e n and e 0 respectively represent the energy standard deviation magnitudes of white noise and the original signal; N is the total average number;

[0037] The different order intrinsic mode functions are used to reflect the different time scales of the original signal. By comparing to obtain the order corresponding to the global minimum of the intrinsic mode energy, the subsequent low-frequency components of this order are used to reconstruct the signal to achieve the purpose of data-driven adaptive denoising; the starting order J s of the reconstruction and the signal after denoising reconstruction are specifically:

[0038]

[0039] In the formula, w represents the w-th time domain value of the intrinsic mode function, and n represents the total length of the time domain of the intrinsic mode function.

[0040] Optionally, performing feature extraction on the denoised data to obtain time-domain and frequency-domain eigenvalue, specifically including:

[0041] Based on the denoised data, performing multi-dimensional feature parameter extraction from the signal time domain and frequency domain respectively. The time-domain signal feature parameters mainly include mean, peak value, root mean square, standard deviation and kurtosis. In order to obtain the frequency-domain information of the received signal, performing fast Fourier transform on the equally spaced time-sampled signal to obtain the frequency spectrum of the signal. Assuming that the amplitude corresponding to the frequency component f(J) in the frequency spectrum is S(J), J = 1, 2,..., m, determining the frequency-domain eigenvalue based on the amplitude. The frequency-domain eigenvalue includes root mean square of frequency amplitude, frequency concentration rate, frequency standard deviation, root mean square frequency and spectral region area.

[0042] Optionally, constructing a multi-layer feedforward artificial neural network based on error backpropagation as a damage degree prediction proxy model, tuning the parameters of the damage degree prediction proxy model, and using the trained damage degree prediction proxy model to identify the time-domain and frequency-domain eigenvalues to obtain an identification result, specifically including:

[0043] Data preprocessing: Through numerical simulation, processing damages of different sizes at different longitudinal positions and transverse positions of the structure to be measured, obtaining damage samples of the structure to be measured and the corresponding received guided wave signals, extracting the multi-dimensional eigenvalues corresponding to each guided wave signal, and classifying the training set and the test set from the eigenvalue vector samples; randomly extracting 70% of the data in the feature set samples as training data, and respectively extracting 15% of the data as the validation set and the test data, and normalizing the eigenvalues according to the following formula:

[0044]

[0045] In the formula: X is the original feature; X nor is the normalized feature; X min is the minimum value in the original feature; X max is the maximum value in the original feature;

[0046] Design and establishment of the model network structure: First, randomly generate the initial weights and thresholds of the network structure from the interval [-1, 1]; determine the number of input layer nodes p according to the dimension of the extracted signal eigenvalues; divide the structural damage degree into q categories, and set q nodes in the output layer; select to construct a BP neural network with m layers for automatic classification of the damage degree. The number of hidden layer nodes is usually determined by the empirical formula:

[0047]

[0048] In the formula, N m represents the number of hidden layer nodes; p and q respectively represent the number of input and output layer nodes; a is an integer between 1 and 10;

[0049] Determine the model training plan;

[0050] Automatic optimization of the model: Each time the BP neural network is trained, initial thresholds and weights are randomly generated, and the genetic algorithm is used to optimize the damaged neural network; First, an initial population is generated according to the weights and thresholds, and the real number coding method is used to encode the genes of the initial population chromosomes. The coding length is determined according to the number of variables to be optimized;

[0051] After the initial chromosome is determined, the continuous evolution of the population is achieved by using chromosome crossover, mutation, and selection operations. Since the gene coding method is real number coding, the real number crossover method of the following formula is used to achieve gene crossover of different chromosomes:

[0052]

[0053] In the formula: α kl is the l-th gene of the k-th chromosome in the population; α rl is the l-th gene of the r-th chromosome; b is a random constant in the interval [0,1]; At the same time, the l-th gene of the r-th chromosome is randomly selected for mutation with a certain probability, and the mutation operation is expressed as:

[0054]

[0055] In the formula: α max and α min correspond to the maximum and minimum values of the real number coding respectively, G max represents the maximum number of evolutions, g is the current iteration number, and d and d' are both random numbers in the interval [0,1].

[0056] Define the reciprocal of the mean square error of the final output of the neural network as the fitness function, and use the fitness function to screen the fitness of each chromosome, retain the chromosomes with high fitness and eliminate the chromosomes with low fitness; The fitness function is expressed as:

[0057]

[0058] In the formula, h u and o u are the actual output and the expected output corresponding to the input sample u respectively, v represents the total number of sample groups, and the smaller the mean square error, the larger the fitness value F(a) of the individual a, and the higher the probability of being selected and inherited to the next generation;

[0059] When the chromosome coding and evolution rules are determined, determine the crossover probability, mutation probability, and the number of generations to terminate evolution;

[0060] Model training and prediction output: Set the expected output of each group of eigenvalue vectors according to the damage category identifier corresponding to the damage signal. After randomly initializing the weights and thresholds of the network structure, input the training feature set for training. During the training process, continuously adjust the weights and thresholds according to the model training scheme. After the optimization training is completed, decode the optimal chromosome that appears to obtain the optimized thresholds and weights of the neural network. Assign the optimized thresholds and weights to the above neural network, and at the same time add a new sample feature set to obtain the classification error change curve and classification confusion matrix diagram of the optimized neural network. Finally, use the trained damage degree prediction proxy model to identify the time-domain and frequency-domain eigenvalue to obtain the identification result.

[0061] Optionally, the determination process of the model training scheme is as follows:

[0062] Adjust the model learning rate according to the training process error curve, set the maximum number of iterations to 200, and terminate the minimum gradient to 1×10 -6 , and set the maximum number of checks to 8;

[0063] Select the scaled conjugate gradient method to update the neural network weight and threshold parameters, and select the Tanh activation function for the nonlinear fitting of the hidden layer, while the output layer selects the Softmax function suitable for nonlinear fitting classification.

[0064] The present invention also provides a non-contact automatic identification system for structural damage, which is used to apply the method as described above, including: a signal excitation system, a signal reception system, and an automatic analysis system; wherein, the signal excitation system includes a signal generator, a first signal amplifier, and an air-coupled transmitting sensor; the signal reception system includes an air-coupled receiving sensor, a second signal amplifier, and a signal oscilloscope; the automatic analysis system includes a computer with a built-in non-contact automatic identification method for structural damage;

[0065] During operation, the signal generator emits a sine excitation signal modulated by a Hanning window, sets the signal intensity emitted by the first signal amplifier gain, and then outputs air-coupled ultrasonic waves after the focusing effect of the air-coupled transmitting sensor; the air-coupled ultrasonic waves form forward-propagating coupled guided waves in the structure to be measured after refraction at the air-structure interface; the air-coupled receiving sensor captures the acoustic wave signal leaked into the air in the structure to be measured, and after being amplified by the second signal amplifier, it is transmitted and stored in the signal oscilloscope; the computer calls the timing monitoring data in the signal oscilloscope and evaluates the degree of structural damage according to the built-in non-contact automatic identification method for structural damage.

[0066] According to the specific embodiments provided by the present invention, the following technical effects of the present invention are disclosed:

[0067] The present invention discloses a non-contact automatic identification method and system for structural damage. The method includes determining the excitation incident angle of the signal relative to the structure to be measured and the sinusoidal excitation signal; determining the tilt angle of the air-coupled receiving sensor at the receiving end according to the excitation incident angle, amplifying and transmitting the received sinusoidal excitation signal to obtain the feedback time-series monitoring data; performing denoising processing on the time-series monitoring data, and performing identification according to the denoised data and the trained damage degree prediction proxy model; the identification result includes whether there is structural damage to the structure to be measured and the degree of structural damage when there is structural damage; the denoising processing uses the ensemble empirical mode decomposition method based on the continuous mean square error criterion. The present invention can solve the problems of poor continuity in traditional ultrasonic guided wave detection and low automation degree in damage degree identification, and improve the classification accuracy of structural damage identification based on air-coupled ultrasonic guided waves. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] Figure 1 It is a schematic structural diagram of the non-contact automatic identification system for structural damage in this embodiment;

[0070] Figure 2 It is a schematic flowchart of the non-contact automatic identification method for structural damage in this embodiment;

[0071] Figure 3 It is a schematic diagram of the correlation between the denoised and reconstructed signal and the original signal in this embodiment; among them, Figure 3 in (a) is a scatter plot of the correlation coefficient distribution between the reconstructed signal and the original signal; Figure 3 in (b) is a schematic diagram of the reconstructed signal with the maximum correlation coefficient; Figure 3 in (c) is a schematic diagram of the reconstructed signal with the minimum correlation coefficient;

[0072] Figure 4 It is a schematic diagram of the comparison of the classification confusion matrix in this embodiment; among them, Figure 4 in (a) is a schematic diagram of the classification confusion matrix corresponding to the noisy feature set for the optimized neural network; Figure 4 in (b) is a diagram of the corresponding classification confusion matrix after denoising by the combined algorithm in this embodiment;

[0073] Figure 5 It is a schematic diagram of the performance of the damage degree classification index under different noise intensities in this embodiment; among them, Figure 5 in (a) is a fitness diagram; Figure 5In (b) is the schematic diagram of the accuracy rate. Specific implementation manners

[0074] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0075] The purpose of the present invention is to provide a non-contact automatic identification method and system for structural damage, which can solve the problems of poor continuity in traditional ultrasonic guided wave detection and low automation degree in damage degree identification.

[0076] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0077] As Figures 1 - 5 shown, the present invention provides a non-contact automatic identification method for structural damage, which is implemented based on a non-contact automatic identification system for structural damage. The non-contact automatic identification system for structural damage mainly consists of a signal excitation system, a signal reception system, and an automatic analysis system. As Figure 1 shown, the signal excitation system consists of a signal generator, a signal amplifier, and an air-coupled sensor; the signal reception system consists of an air-coupled sensor, a signal amplifier, and a signal oscilloscope; the automatic analysis system consists of a computer and an embedded automatic damage degree identification algorithm. The signal generator is used to emit a sine excitation signal modulated by a Hanning window, the gain of the signal amplifier is set to adjust the intensity of the emitted signal, and then the air-coupled ultrasonic wave is output after the focusing effect of the air-coupled ultrasonic sensor; the air-coupled ultrasonic wave forms a forward-propagating coupled guided wave in the structure to be measured after refraction at the air-structure interface; the acoustic wave signal leaked into the air in the structure is captured by the air-coupled sensor in the signal reception system, and after being amplified by the signal amplifier, it is transmitted and stored in the signal oscilloscope; the signal analysis system calls the time-series monitoring data in the signal oscilloscope and evaluates the structural damage degree according to the embedded automatic identification and classification algorithm. The specific process is as follows:

[0078] Step 1: Determine the excitation scheme of the signal generator: The load excited by the air-coupled guided wave detection technology does not directly act on the structure, but first propagates in the air medium in the form of acoustic waves, and after reaching the air-structure interface, it refracts and enters the structure to excite the guided wave. In order to excite the guided wave with the maximum energy propagating forward, it is necessary to ensure that the acoustic wave refraction angle is 90°. According to Snell's law, the excitation incident angle θ is determined:

[0079]

[0080] where c 空气 represents the acoustic wave velocity in the air medium, and c 结构 represents the Rayleigh guided wave velocity in the structural medium, and θ represents the excitation incident angle of the air-coupled ultrasonic sensor.

[0081] Select a Hanning window modulated sine signal with a small sidelobe amplitude and concentrated energy at the center frequency as the excitation signal of the signal generator, and its expression is:

[0082]

[0083] where x(t) represents the Hanning window modulated sine signal; f c represents the center frequency of the modulation signal; T represents the number of periods of the modulation signal.

[0084] It can be seen from Equation (2) that the excitation signal is jointly determined by the center frequency and the number of periods. The optimal excitation signal parameters are determined according to the multimodal characteristics and dispersion effects of the structure involved. Connect the signal output by the signal generator to the input end of the signal amplifier, and after signal gain amplification, output. The output end of the signal amplifier outputs air-coupled ultrasonic waves after the focusing effect of the air-coupled ultrasonic sensor. The air-coupled ultrasonic waves form forward-propagating coupled guided waves in the structure to be measured after refraction at the air-structure interface.

[0085] Step 2: Determine the signal receiving system, and capture the acoustic wave signal leaked from the structure to the air by the air-coupled sensor in the signal receiving system; in order to receive the acoustic wave with the maximum energy, determine the tilt angle of the air-coupled sensor at the receiving end according to Equation (1), and after signal amplifier gain, transmit and store it in the signal oscilloscope for real-time download and analysis by the automatic analysis system.

[0086] Step 3: The automatic analysis system calls the timing monitoring data in the signal oscilloscope and evaluates the structural damage degree according to the embedded automatic recognition algorithm. It is characterized in that the automatic denoising of the guided wave signal with a low signal-to-noise ratio is realized based on the adaptive denoising algorithm, and the intelligent optimization algorithm and the prediction classification algorithm are used for automatic extraction and classification of damage information, which is applicable to the automatic recognition of the structural damage degree based on air-coupled ultrasonic guided waves. The specific implementation process is as follows:

[0087] (1) Due to the difference in acoustic impedance between the two different media of the structure and the air, the signal-to-noise ratio intensity of the acoustic wave signal captured by the signal receiving system is theoretically low. In order to reduce the interference of noise on the signal and realize signal automatic denoising and damage degree recognition, an adaptive denoising method based on energy mutation mode order reconstruction of the signal is proposed for the narrowband characteristics of the guided wave excitation signal. Specifically, for an original signal x(t), it is decomposed step by step according to the ensemble empirical mode decomposition method to obtain multiple order intrinsic mode functions:

[0088] The local maxima and minima of the signal \(x(t)\) are obtained separately. Cubic spline interpolation fitting is performed on the envelopes of the local maxima and minima. The fitted envelopes are summed and averaged, and the average value is \(m(t)\). The average value \(m(t)\) is subtracted from the original signal \(x(t)\) to obtain \(h(t)=x(t)-m(t)\).

[0089] It is judged whether the intermediate signal \(h(t)\) satisfies the two conditions of the intrinsic mode function \(F(t)\): (a) The total number of extreme points of the signal is equal to or differs by only one from the total number of zero-crossing points; (b) The average value of the upper envelope line fitted by the local maxima of the signal and the lower envelope line fitted by the local minima should be zero. This limiting condition can solve the problem of instantaneous frequency fluctuation caused by waveform asymmetry. If \(h(t)\) satisfies these two conditions, then \(h(t)\) is an intrinsic mode component; otherwise, this intermediate signal \(h(t)\) is used to replace \(x(t)\), and the above steps are repeatedly calculated until \(h(t)\) satisfies the conditions.

[0090] After obtaining the first intrinsic mode \(F\) 1 (t), subtract \(F\) 1 (t) from the original signal \(x(t)\) to get \(x\) 1 (t)=x(t)-F 1 (t). Let \(x\) 1 (t) be the new original signal, and repeat the above steps to obtain \(F\) 2 (t). According to the above steps, the intrinsic mode functions of each order are obtained in turn. When the intrinsic mode function is a monotonic function, the above signal mode decomposition can be terminated. Finally, the original signal is decomposed into a series of intrinsic mode functions \(F\) i (t) and the residual term \(R(t)\), as shown in Equation (3):

[0091]

[0092] In the formula, \(i\) represents the order of the intrinsic mode function, and \(M\) represents the total number of orders of the decomposed intrinsic mode functions. The phenomenon of modal aliasing of different scales is eliminated through the collection and averaging steps of the modal decomposition process. The specific steps are as follows:

[0093] Add different white noises to the original signal multiple times to generate multiple different noisy signals to be processed:

[0094] x j (t)=x(t)+n j (t) (4)

[0095] Among them, \(n\) j (t) is the white noise sequence added for the \(j\)th time, \(j = 1, 2,\cdots\). Empirical mode decomposition is performed on different noisy signals respectively, and multiple groups of intrinsic mode components \(F\) ij (t), add and average the components of the corresponding orders in different groups of F i j in (t) as the final result F of the eigenmode component of this order i (t):

[0096]

[0097] where F i j (t) represents the mixed signal x with white noise added for the jth time j the ith-order eigenmode component obtained after the above decomposition of (t), and N represents the total average number of times

[0098] By the set and average steps, the mode aliasing caused by noise is eliminated. When N is large enough, taking the set and average of F i j (t) can cancel out the white noise components with each other, and then obtain the ideal intrinsic mode function F i (t). The relationship between the average number of times and the decomposition error in the ensemble empirical mode decomposition method can be determined according to Equation (6):

[0099]

[0100] In the formula, e represents the ensemble empirical mode decomposition error, e n and e 0 represent the energy standard deviation magnitudes of white noise and the original signal respectively; N is the total average number of times, and this parameter can be adjusted according to the decomposition effect

[0101] The different-order intrinsic mode functions reflect different time scales of the original signal. The first-order intrinsic mode function F 1 (t) component carries a large amount of high-frequency information. As the order increases, the low-frequency components in the signal gradually appear. Therefore, there must be a certain order of components such that the energy of the subsequent intrinsic mode function signal is greater than the energy of random noise. By comparing to obtain the order corresponding to the minimum global intrinsic mode energy, and using the subsequent low-frequency components of this order to reconstruct the signal to achieve the purpose of data-driven adaptive denoising. The starting order J of reconstruction s and the signal after denoising reconstruction see Equations (7) and (8).

[0102]

[0103] In the formula, w represents the wth time-domain value of the intrinsic mode function, and n represents the total length of the time domain of the intrinsic mode function

[0104] 2) Information extraction and analysis of the denoised signal: Multidimensional feature parameters are extracted from both the time domain and frequency domain of the signal. The time domain signal feature parameters mainly include mean, peak value, root mean square, standard deviation, kurtosis, etc. In order to obtain the frequency domain information of the received signal, a fast Fourier transform is performed on the equally spaced time-sampled signal to obtain the frequency spectrum of the signal. Assuming that the amplitude corresponding to the frequency component f(J) in the frequency spectrum is S(J), J = 1, 2,..., m, the frequency domain eigenvalues such as root mean square of frequency amplitude, frequency concentration rate, frequency standard deviation, root mean square frequency, and spectral region area are further obtained accordingly.

[0105] 3) Automatic identification and classification of structural damage degree: A multi-layer feedforward artificial neural network based on error backpropagation (BP) is constructed as a damage degree prediction proxy model. The model includes four parts: data preprocessing, network structure design and establishment, model optimization, and model training and prediction output. The specific construction steps are summarized as follows:

[0106] (a) Data preprocessing

[0107] By numerical simulation, damages of different sizes are processed at different longitudinal and transverse positions of the structure to be measured, and the received guided wave signals corresponding to a sufficient amount of damage samples of the structure to be measured are obtained. The multidimensional eigenvalues corresponding to each guided wave signal are extracted, and the training set and test set are classified from the eigenvalue vector samples. 70% of the data is randomly selected from the feature set samples as training data, 15% of the data is respectively selected as the validation set and test data, and the eigenvalues are normalized according to Equation (9).

[0108]

[0109] where: X is the original feature; X nor is the normalized feature; X min is the minimum value in the original feature; X max is the maximum value in the original feature.

[0110] (b) Design and establishment of the model network structure

[0111] First, the initial weights and thresholds of the network structure are randomly generated from the interval [-1, 1]. The number of input layer nodes p is determined according to the dimension of the extracted signal eigenvalues. The structural damage degree is divided into q categories, and q nodes are set in the output layer. A BP neural network with m layers is selected to construct for automatic classification of damage degree. The number of hidden layer nodes is usually determined by the empirical formula:

[0112]

[0113] where, N m represents the number of hidden layer nodes; p and q respectively represent the number of input and output layer nodes; a is an integer between 1 and 10.

[0114] Adjust the model learning rate according to the training process error curve. To prevent the training process from taking too long, set the maximum number of iterations to 200, and the termination minimum gradient to 1×10 -6 . To prevent overfitting during the training learning process, set the maximum number of checks to 8, that is, if the training error does not decrease for 8 consecutive times, the training will be terminated.

[0115] Select the scaled conjugate gradient method with fast convergence speed and less memory occupation to update the neural network weight and threshold parameters. Referring to the characteristics of each activation function, select the Tanh activation function with faster convergence speed for the non-linear fitting of the hidden layer, while the output layer selects the Softmax function suitable for non-linear fitting classification.

[0116] (c) Automatic optimization of the model

[0117] Each time the BP neural network is trained, initial thresholds and weights are randomly generated. Different initial parameters lead to different final convergence results. To improve the prediction stability and accuracy of the network model and avoid manual repeated debugging work, use the genetic algorithm to optimize the damaged neural network. First, generate an initial population according to the optimization variables (weights and thresholds). Since the decision variables to be optimized (weights and thresholds) are real numbers and continuous variables, a real number coding method is used to encode the genes of the initial population chromosomes, and the coding length depends on the number of variables to be optimized.

[0118] After the initial chromosome is determined, operations such as chromosome crossover, mutation, and selection are used to realize the continuous evolution of the population. Since the gene coding method is real number coding, a real number crossover method as shown in Equation (11) is used to realize the gene crossover of different chromosomes:

[0119]

[0120] In the formula: α kl is the l-th gene of the k-th chromosome in the population; α rl is the l-th gene of the r-th chromosome; b is a random constant in the interval [0,1]. At the same time, to avoid the search falling into a local area, the l-th gene of the r-th chromosome is randomly selected for mutation with a certain probability, and the mutation operation can be expressed as:

[0121]

[0122] In the formula: α max and α min correspond to the maximum and minimum values of the real number coding respectively, G max represents the maximum number of evolutions, g is the current iteration number, and d and d' are both random numbers in the interval [0,1].

[0123] Define the reciprocal of the mean square error of the final output of the neural network (Equation (14)) as the fitness function, evaluate the fitness of each chromosome, retain the chromosomes with high fitness, and eliminate the chromosomes with low fitness.

[0124]

[0125] Where h u and o u are the actual output and the desired output corresponding to the input sample u respectively, v represents the total number of sample groups. The smaller the mean square error, the larger the fitness value F(a) of the individual a, and the higher the probability of being selected and inherited to the next generation.

[0126] After determining the chromosome coding and evolution rules, under the principle of taking into account the optimization ability and efficiency of the algorithm, parameters such as the crossover probability, mutation probability, and termination evolution algebra are determined through trial calculations.

[0127] (d) Model training and prediction output

[0128] Set the desired output of each group of eigenvalue vectors according to the damage class identifier corresponding to the damage signal. After randomly initializing the weights and thresholds of the network structure, input the training feature set for training, and continuously adjust the weights and thresholds according to the optimization algorithm during the training process. After the optimization training is completed, decode the obtained optimal chromosome to obtain the optimized thresholds and weights of the neural network, assign the optimized thresholds and weights to the above neural network, and at the same time add a new sample feature set to obtain the classification error change curve and classification confusion matrix diagram of the optimized neural network.

[0129] Example:

[0130] Taking the rail in the railway traffic infrastructure system as an example, the following specific implementation process is provided.

[0131] 1) Determine the monitoring scheme

[0132] Theoretically, in order to be able to excite the Rayleigh guided wave with the maximum energy propagating forward, it is necessary to ensure that the acoustic wave refraction angle is 90°, and according to Snell's law, that is, to satisfy the equation:

[0133]

[0134] The Rayleigh wave velocity of the material can be estimated by the following formula:

[0135]

[0136] Substitute the Poisson's ratio υ and the shear wave velocity C T of the rail material, and the theoretical Rayleigh wave velocity C R of the rail can be obtained as 2946 m / s. Given that the acoustic wave velocity in air is 343 m / s, the acoustic wave incident angle θ 1It is equal to 6.6°. Similarly, in order to receive the sound wave with the maximum energy, the receiving sensor needs to be at an angle of 6.6° with the vertical line of the interface.

[0137] As can be seen from Equation (2), the excitation signal is jointly determined by the center frequency and the number of cycles. The optimal excitation signal parameters are determined according to the multimodal characteristics and dispersion effects of the structure involved.

[0138] 2) Numerical simulation of rail damage monitoring based on the monitoring scheme

[0139] Based on the general commercial finite element software, the whole process of rail non-destructive testing based on the above monitoring scheme is simulated. In the model, the rail structure is still simulated by three-dimensional 8-node reduced integration elements. The air domain surrounding the rail is simulated by defining an acoustic medium, and the air domain elements are simulated by acoustic elements. In order to simulate an infinite air domain, non-reflective boundary conditions are defined around the air column to avoid the interference of acoustic wave reflection at the side boundary of the air column in the model. According to the acoustic motion equation, the propagation of sound waves in the air is simulated, and the displacement field of the rail structure in the model is coupled with the pressure field of the air domain by using the coupled motion equation. The accuracy of the acoustic-solid coupling model is verified by the arrival time of the wave packet signal and the sound pressure amplitude.

[0140] In view of the fact that it is difficult to obtain a large number of different rail damage samples in reality, different degrees of rail damage are processed in the established finite element model. By using the full-process simulation function of air-coupled guided waves in the finite element model, a sufficient amount of signal samples containing different rail damage information are obtained. Considering the particularity of the damage at the rail bottom and the necessity of detection, different degrees of damage are generated at different positions at the bottom of the rail. After determining the longitudinal position, rail defects with different damage degrees are generated at different transverse positions respectively. A total of 1870 different damage models are processed. A new command-type bat file is created, and the Interactive command is written in the bat file to achieve sequential automatic submission of calculations for different damage models. By writing the CPU call command, efficient multi-core calculation is achieved, avoiding the problems of manual submission of operations and waiting for calculation time-consuming.

[0141] 3) Analysis of damage guided wave signals and identification and classification of damage degrees (as shown in Table 1 and Table 2)

[0142] To further study the possibility and accuracy of automatic identification and classification of damage under the action of noise pollution, high-intensity Gaussian white noise randomly distributed in the interval of signal-to-noise ratio from 0 dB to -6 dB (the noise energy is 1 to 4 times the energy of the reference signal) is added. In order to comprehensively cover the damage information in the received signal, multi-dimensional characteristic parameters are extracted from the time domain and frequency domain of the noise signal respectively. The statistical results of its main characteristic parameters are shown in Table 1.

[0143] Table 1 Statistical characteristic parameters of the sampled signal in the time domain

[0144]

[0145] Table 2 Sampling Signal Frequency Domain Statistical Feature Parameters

[0146]

[0147]

[0148] Considering that it is difficult to extract effective time-domain feature information from noise interference signals, the noise interference signals are analyzed in the frequency domain, the frequency domain feature parameters in the corresponding frequency range are extracted and the feature parameters are normalized. Neural network training and genetic algorithm optimization are carried out based on the multi-dimensional frequency domain eigenvalue vectors of 1000 groups of damage guided wave signals. During the whole optimization process, the reciprocal of the average fitness of the population is always in a fluctuating state and there is no obvious decrease, while the reciprocal of the optimal fitness only decreases slightly in the early stage of evolution and then remains unchanged. It can be seen that the quality of the feature set information under noise interference is poor, resulting in insufficient optimization potential of the damage classification neural network. Draw an intuitive classification confusion matrix diagram, as Figure 4 shown in (a), its overall classification accuracy is 77.3%, while the classification accuracy of type II damage is only about 50%.

[0149] Use the combined algorithm to realize the automatic reconstruction of the target signal. As Figure 3 shown in (a) of, the distribution interval of the reconstruction correlation coefficients of 1870 groups of random noise interference signals is 0.93 - 0.99, mainly distributed between 0.97 - 0.99, indicating that the quality of the denoised signals obtained by the above method is relatively high. Among them, the reconstructions with the maximum and minimum correlation coefficients are shown in Figure 3 (b) of and Figure 3 (c) of respectively, and the distinguishability of the effective features of the signals is enhanced. According to the eigenvalue expressions shown in Tables 1 and 2, the eigenvalues of the reconstructed main wave packet are extracted and input into the neural network optimized by the genetic algorithm for training and prediction. The classification confusion matrix shows that the overall classification accuracy of the samples reaches 84.9%. Compared with Figure 4 the classification result without denoising processing by the combined algorithm shown in (a) of, the overall accuracy is increased by 7.6 percentage points. More importantly, the recognition rate of type II damage in the overall samples is increased from 50% to 64%.

[0150] The results show that the combined algorithm of the present invention can effectively improve the recognizability of the feature set of noisy signals and increase the classification accuracy of damage samples. This algorithm can automatically realize multiple functions such as adaptive denoising of signals, automatic feature extraction, intelligent optimization of neural networks and automatic classification, and also has certain application prospects in the recognition of the damage degree of other structures.

[0151] 4) Robustness Analysis

[0152] For the classification algorithm of the present invention, robustness represents its ability to resist data perturbation and noise interference in classification results. Figure 5 The following shows the performance indicators of automatic damage classification using the above combined algorithm under different noise intensities. As the noise intensity increases, the population fitness shows a decreasing trend, but generally changes little. The classification accuracy and mean square error have small fluctuations and are still in a concentrated and stable state as a whole. This proves that the classification calculation results of the combined algorithm are stable and reliable, and have good robustness against different intensities of noise interference.

[0153] 5) Conclusion

[0154] The above results show that the structural damage identification system and method based on air-coupled guided waves in this embodiment can be used for non-contact continuous monitoring of internal structural damage and has excellent effects: through the non-contact automatic identification algorithm for structural damage based on air-coupled guided waves, multiple functions such as adaptive denoising of signals, automatic feature extraction, intelligent optimization of prediction models, and automatic identification of damage degree can be automatically realized, which can effectively improve the recognizability of the feature set of noisy signals and the identification accuracy of damage degree. The robustness test results show that the algorithm has the ability to resist data perturbation, noise intensity interference, and the influence of parameter fluctuations, and has good application prospects.

[0155] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other.

[0156] Specific examples are used in this article to elaborate on the principles and implementation methods of the present invention. The descriptions of the above embodiments are only used to help understand the core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A non-contact automatic identification method for structural damage, characterized in that: include: determining an excitation incident angle of the signal relative to the structure to be measured and a sinusoidal excitation signal; Determine the tilt angle of the air-coupled receiving sensor at the receiving end according to the excitation incident angle, and amplify and transmit the received sinusoidal excitation signal to obtain feedback timing monitoring data; De-noising the time series monitoring data, and performing identification based on the de-noised data and the trained damage degree prediction agent model; the identification result includes whether the structure to be tested has structural damage and the degree of structural damage when structural damage exists; The denoising process adopts an ensemble empirical mode decomposition method based on a continuous mean square error criterion.

2. The non-contact automatic identification method of structural damage according to claim 1, characterized in that: The method for determining the excitation incident angle and the sinusoidal excitation signal is: Set the sound wave refraction angle to 90° and determine the excitation incident angle θ according to Snell's law: In the formula, c 空气 represents the speed of sound waves in air, c 结构 represents the Rayleigh guided wave velocity in the structure medium, θ represents the excitation incident angle of the air-coupled ultrasonic sensor; The sine signal modulated by the Hanning window is selected as the excitation signal of the signal generator, which is expressed as: Where x(t) represents the sinusoidal signal modulated by the Hanning window; f c represents the center frequency of the modulation signal; T represents the number of cycles of the modulation signal.

3. The non-contact automatic identification method of structural damage according to claim 1, characterized in that: The denoising process is performed on the time series monitoring data, and recognition is performed based on the denoised data and the trained damage degree prediction agent model to obtain a recognition result, which specifically includes: De-noising the time series monitoring data using an ensemble empirical mode decomposition method based on a continuous mean square error criterion to obtain de-noised data; Performing feature extraction on the denoised data to obtain time domain and frequency domain feature values; A multi-layer feedforward artificial neural network based on error back propagation is constructed as a damage degree prediction agent model, parameters of the damage degree prediction agent model are tuned, and the time domain and frequency domain eigenvalues ​​are identified using the trained damage degree prediction agent model to obtain an identification result.

4. The non-contact automatic identification method of structural damage according to claim 3 is characterized in that: The denoising process of the time series monitoring data using the ensemble empirical mode decomposition method based on the continuous mean square error criterion to obtain the denoised data specifically includes: Assume that the original signal in the time series monitoring data is x(t), and perform step-by-step decomposition according to the ensemble empirical mode decomposition method to obtain multi-order intrinsic mode functions: The local maximum and minimum values ​​of the original signal x(t) are obtained respectively, and the envelopes of the local maximum and minimum values ​​are fitted by cubic spline interpolation. The fitted envelopes are summed and averaged to obtain the average value m(t), and the average value m(t) is subtracted from the original signal x(t) to obtain the intermediate signal h(t)=x(t)-m(t); Determine whether the intermediate signal h(t) satisfies two conditions of the intrinsic mode function F(t): (a) the total number of signal extreme value points is equal to the total number of zero-crossing points or differs by only one; (b) the average value of the upper envelope fitted by the local maximum value of the signal and the lower envelope fitted by the local minimum value of the signal should be zero; if h(t) satisfies the two conditions, then h(t) is an intrinsic mode component; otherwise, use the intermediate signal h(t) to replace x(t), and repeat the calculation of the previous step until the recalculated intermediate signal h(t) satisfies the conditions; After obtaining the first eigenmode F1(t), subtract F1(t) from the original signal x(t) to obtain x1(t)=x(t)-F1(t). Let x1(t) be the new original signal, and re-execute the operation steps to obtain F2(t). According to the above steps, the eigenmode functions of each order are obtained in turn. When the eigenmode function is a monotonic function, the signal modal decomposition is terminated. Finally, the original signal is decomposed into a series of eigenmode functions F i (t) and the residual term R(t), expressed as: Where i represents the order of the intrinsic mode function, and M represents the total order of the decomposed intrinsic mode function; The aliasing of modes of different scales is eliminated through the aggregation and averaging steps of the modal decomposition process. The specific steps are as follows: Add different white noises to the original signal multiple times to generate multiple different noisy signals to be processed: j (t) = x(t) + n j (t); Among them, n j (t) is the jth added white noise sequence, j = 1, 2, ..., perform empirical mode decomposition on different noisy signals to obtain multiple groups of intrinsic mode components F i j (t), different groups F i j The corresponding order components in (t) are added and averaged to obtain the final result F of the eigenmode component of this order. i (t): Among them, F i j (t) represents the mixed signal x with white noise added for the jth time j (t) The i-th order eigenmode component obtained after decomposition, N represents the total average number; The modal aliasing caused by noise is eliminated by the collection and averaging steps. When N reaches the set threshold, F i j (t) Collect and average the white noise components to cancel each other out and obtain the ideal intrinsic mode function F i (t); In the integrated modal decomposition method, the relationship between the average number and the decomposition error is determined by the following formula: In the formula, e represents the integrated empirical mode decomposition error, e n and e0 represent the energy standard deviation of white noise and original signal respectively; N is the total average number; Different orders of eigenmode functions are used to reflect the different time scales of the original signal. The order corresponding to the minimum value of the global eigenmode modal energy is obtained by comparison. The signal is reconstructed using the subsequent low-frequency components of this order to achieve the purpose of data-driven adaptive denoising. The reconstruction starting order J s And the denoised and reconstructed signal Specifically: Where w represents the wth time domain value of the intrinsic mode function, and n represents the total time domain length of the intrinsic mode function.

5. The non-contact automatic identification method of structural damage according to claim 3, characterized in that: The feature extraction of the denoised data to obtain time domain and frequency domain feature values ​​specifically includes: Based on the denoised data, multidimensional feature parameters are extracted from the signal time domain and frequency domain respectively, and the time domain signal feature parameters mainly include mean, peak, root mean square, standard deviation and kurtosis; in order to obtain the frequency domain information of the received signal, the signal is sampled at equal intervals and fast Fourier transform is performed to obtain the frequency spectrum of the signal. It is assumed that the amplitude corresponding to the frequency component f(J) in the spectrum is S(J), J=1,2,...,m, and the frequency domain feature value is determined based on the amplitude. The frequency domain feature values ​​include the root mean square of the frequency amplitude, the frequency concentration rate, the frequency standard deviation, the root mean square frequency and the area of ​​the spectrum region.

6. The non-contact automatic identification method of structural damage according to claim 3, characterized in that: The multi-layer feedforward artificial neural network based on error back propagation is constructed as a damage degree prediction agent model, the parameters of the damage degree prediction agent model are tuned, and the time domain and frequency domain feature values ​​are identified using the trained damage degree prediction agent model to obtain the identification results, which specifically include: Data pre-processing: Through numerical simulation, damages of different sizes are processed at different longitudinal and transverse positions of the structure to be tested, and the damage samples of the structure to be tested and the corresponding received waveguide signals are obtained. The multi-dimensional eigenvalues ​​corresponding to each waveguide signal are extracted, and the training set and the test set are classified from the eigenvalue vector samples; 70% of the data are randomly selected from the feature set samples as training data, and 15% of the data are respectively selected as the verification set and the test data, and the eigenvalues ​​are normalized according to the following formula: Where: X is the original feature; X nor is the homogenized feature; X min is the minimum value of the original feature; X max is the maximum value among the original features; Design and establishment of the model network structure: First, randomly generate the initial weights and thresholds of the network structure from the interval [-1,1]; determine the number of input layer nodes p according to the dimension of the extracted signal eigenvalue; divide the structural damage degree into q categories, and set q nodes in the output layer; choose to build an m-layer BP neural network for automatic damage degree classification, and the number of hidden layer nodes is usually determined by an empirical formula: Where N m represents the number of hidden layer nodes; p and q represent the number of input and output layer nodes respectively; a is an integer between 1 and 10; Determine the model training plan; Automatic optimization of the model: Each time the BP neural network is trained, the initial threshold and weight are randomly generated, and the damaged neural network is optimized using a genetic algorithm. First, the initial population is generated based on the weight and threshold, and the real number coding method is used to genetically encode the chromosomes of the initial population. The coding length is determined according to the number of variables to be optimized. After the initial chromosome is determined, chromosome crossover, mutation and selection operations are used to achieve continuous evolution of the population. Since the gene encoding method is real number encoding, the following real number crossover method is used to achieve gene crossover of different chromosomes: Where: α kl is the lth gene of the kth chromosome in the population; α rl is the lth gene of the rth chromosome; b is a random constant in the interval [0,1]; at the same time, the lth gene of the rth chromosome is randomly selected for mutation according to a certain probability, and the mutation operation is expressed as: Where: α max and α min Corresponding to the maximum and minimum values ​​of real number coding, G max represents the maximum number of evolutions, g is the current number of iterations, and d and d' are both random numbers in the interval [0,1]. The inverse of the mean square error of the final output of the neural network is defined as the fitness function, and the fitness function is used to screen the fitness of each chromosome, retaining chromosomes with high fitness and eliminating chromosomes with low fitness; the fitness function is expressed as: In the formula, h u and u are the actual output and expected output corresponding to the input sample u, respectively. v represents the total number of sample groups. The smaller the mean square error, the larger the fitness value F(a) of individual a is, and the higher the probability of being selected and inherited to the next generation. When the chromosome coding and evolutionary rules are determined, the crossover probability, mutation probability and termination evolutionary generation are determined; Model training and prediction output: The expected output of each group of eigenvalue vectors is set according to the damage category identifier corresponding to the damage signal. After randomly initializing the weights and thresholds of the network structure, the training feature set is input for training. During the training process, the weights and thresholds are continuously adjusted according to the model training plan. After the optimization training is completed, the optimal chromosome that appears is decoded to obtain the optimized threshold and weight of the neural network. The optimized threshold and weight are assigned to the above-mentioned neural network. At the same time, a new sample feature set is added to obtain the classification error change curve of the optimized neural network and the classification confusion matrix diagram. Finally, the trained damage degree prediction agent model is used to identify the time domain and frequency domain eigenvalues ​​to obtain the recognition result.

7. The non-contact automatic identification method of structural damage according to claim 6, characterized in that: The process of determining the model training scheme is as follows: Adjust the model learning rate according to the error curve of the training process, set the maximum number of iterations to 200, and the minimum termination gradient to 1×10 -6 , and set the maximum number of checks to 8; The proportional conjugate gradient method is selected to update the neural network weights and threshold parameters, and the Tanh activation function is selected for the nonlinear fitting of the hidden layer, while the Softmax function suitable for nonlinear fitting classification is selected for the output layer.

8. A non-contact automatic identification system for structural damage, used for applying the method according to any one of claims 1 to 7, characterized in that: include: A signal excitation system, a signal receiving system and an automatic analysis system; wherein the signal excitation system comprises a signal generator, a first signal amplifier and an air-coupled transmitting sensor; the signal receiving system comprises an air-coupled receiving sensor, a second signal amplifier and a signal oscilloscope; the automatic analysis system comprises a computer with a built-in non-contact automatic identification method for structural damage; During operation, a signal generator transmits a sinusoidal excitation signal modulated by a Hanning window, and the signal strength emitted by the first signal amplifier gain is set, and then air-coupled ultrasonic waves are output after the focusing effect of the air-coupled transmitting sensor; the air-coupled ultrasonic waves are refracted through the air-structure interface to form a coupled guided wave that propagates forward in the structure to be tested; the air-coupled receiving sensor captures the sound wave signal leaked into the air from the structure to be tested, and after being amplified by the second signal amplifier, it is transmitted and stored in the signal oscilloscope; the computer calls the timing monitoring data in the signal oscilloscope, and evaluates the degree of structural damage according to the built-in non-contact automatic identification method of structural damage.

Images